-ProCapNet adapts the BPNet architecture. It reads 2,114 bp of one-hot encoded -sequence and outputs both a base-resolution profile over the central 1,000 bp, -covering both strands as a single softmax so the model can learn strand -asymmetry, and a scalar giving the log total number of initiation events in that -window. Training used ENCODE PRO-cap read alignments in which the first read of -each pair was discarded and only the single 5'-most base of the second read was -kept, merged across replicates and kept separate by strand. Each model was -trained on all PRO-cap peaks in its cell line plus randomly sampled -DNase-hypersensitive sites from the same cell line at a 7:1 peak to background -ratio, with 7-fold cross-validation split by chromosome. Bases that are not -uniquely mappable by 36-mer reads were given zero loss weight during training. -Full details are in Cochran et al., 2024. +ProCapNet adapts the BPNet architecture: it reads 2,114 bp of sequence and +predicts a base-resolution initiation profile over the central 1,000 bp on both +strands, together with the total number of initiation events in that window. One +model was trained per cell line on ENCODE PRO-cap data, using all PRO-cap peaks +in that cell line plus sampled DNase-hypersensitive sites as background, with +7-fold cross-validation split by chromosome. Full details are in Cochran +et al., 2024.
--Genome-wide predictions were generated by the Kundaje lab by applying the model -to every 2,114 bp window at a stride of 250 bp, so each base is the average of -four overlapping predictions, then averaging across the seven cross-validation -models and across the forward and reverse-complemented sequence. On hg38 no -prediction was made where most of a window was unresolved (N) in the reference. +The Kundaje lab generated the genome-wide predictions by applying each model in +2,114 bp windows at a stride of 250 bp, then averaging across the seven +cross-validation models and across the forward and reverse-complemented +sequence. On hg38 no prediction was made where most of a window was unresolved +(N) in the reference. Contribution scores were computed with DeepSHAP, which +contrasts the model's output on the real sequence against its output on 25 +dinucleotide shuffles of it, and were computed only at MANE Select transcription +start sites.
--Scores were computed with DeepSHAP, which estimates each base's contribution by -contrasting the model's output on the real sequence against its output on a set -of reference sequences, here 25 dinucleotide shuffles of the sequence being -scored. Because DeepSHAP needs a single scalar to explain, the base-resolution -profile output was summarized by mean-normalizing the pre-softmax logits and -taking their dot product with the post-softmax profile, which weights each -base's logit by its predicted probability of being used and sums over the -1,000 bp output window and both strands. This is the profile or TSS-positioning -task; ProCapNet can also produce scores for its read-count task, which are not -included in this track. Each scored sequence was run through all seven cross-validation -models and in both orientations, and the scores averaged. -
- -The ProCapNet model implementation is at kundajelab/ProCapNet. The steps that turned the published files into these tracks are recorded in doc/$db/transcriptionStart.txt, the scripts they run are in makeDb/outside/proCapNet, and the track configuration is in trackDb/human/$db/transcriptionStart.ra.